AI That Earns Its Place: Building Network Intelligence for Millions of Homes

Nobody wants an AI feature. They want their WiFi to work

Consumer AI in a home network has a very short leash. The user does not open the app to be delighted; they open it because something is slow in the bedroom, or a child is online at midnight, or a device they don't recognise has joined the network. If intelligence doesn't resolve one of those moments, it's decoration.

So the bar I hold features to is simple: does the model prevent a support call, or does it make the app worth opening on a good day? Smart WiFi, the multi-module ecosystem we launched, is built around that test. Parental controls, group WiFi management, network security, and real-time device analytics are four different reasons to open an app that people used to open only when they were angry.

IoT gives you signals, not answers

A home network produces an enormous amount of telemetry. The temptation is to surface it, because it's there and it looks impressive on a dashboard. Real-time device analytics are only useful when they answer a question the subscriber was already asking.

Signal quality per device is data. "Your streaming stick has a weak connection, and moving the router helps" is a product. The work between those two sentences is where most of the effort in an IoT platform actually goes, and it is mostly product judgement rather than modelling.

App-led growth changes what the roadmap is for

Turning a utility app into a growth surface means every module has to earn attention twice: once as a genuinely useful capability, and once as a reason to come back. That's why we integrated behavioural signals into lead generation and retention workflows, and why adoption per module matters more to me than a headline install count.

An app-led platform lives or dies on the second visit. Installs are a marketing outcome. Return visits are a product one.

LLMs belong where the load is, not where the demo is

The most valuable place we're applying LLMs is customer support automation and agentic workflows, working closely with engineering. That's not the flashiest application, but support is where the volume, the cost, and the frustration all concentrate at once.

An agent that resolves a genuine connectivity issue without a human is worth more than any conversational feature bolted onto a home page. And because the failure mode of a wrong answer in support is a second, angrier contact, it forces the discipline that AI features need anyway: know when the model should hand over, and make that handover fast.

The rule I keep coming back to

Intelligence should reduce the number of times a customer has to think about their broadband. Every model we ship gets measured against that, not against how advanced it sounds in a roadmap review.